Understand your care. Act with confidence.

We help patients not just read their records — but actually know what to do.

Next.js React TypeScript Tailwind CSS

Navis is a patient-facing web app for making sense of discharge papers and visit summaries. Upload a PDF, talk with an AI assistant grounded in your document, watch the PDF scroll and highlight as the assistant responds, and export critical next steps, medication lists (PDF), calendar reminders (.ics), and nearby pharmacies (U.S. ZIP → Google Maps).

Your file stays in the browser session for this app—there is no Navis server that stores your PDF after you leave.

• At a glance
• What it does
• Features
• Tech stack & versions
• Architecture
• Privacy & data
• Requirements
• Environment variables
• Getting started
• Deploying (Vercel)
• npm scripts
• Project structure
• Browser permissions
• API keys & security
• Medical disclaimer
• Roadmap
• License

1. Landing — Upload a PDF. First-time upload can show a privacy brief; the file is held in memory (blob URL) for the session.
2. Document workspace — Split layout: PDF viewer (left on large screens, top on small) and tools in the side column.
3. Live voice (optional) — With a Google API key, start Talk about this document: Gemini Live reads your document context, answers questions, and the UI parses the assistant’s transcript to scroll and highlight on the PDF.
4. Extractions — With Google and/or Groq keys, Critical next steps and medication schedule JSON extractions; downloads for PDF and ICS where applicable.
5. Pharmacies — Valid 5-digit U.S. ZIP opens a Google Maps search URL (no Maps API key in-app).

• .pdf from the file picker on the home page or Re-upload in the header on /document.
• No server-side document database in this repository—the PDF is a blob URL in the browser.

• PDF.js renders pages and a text layer.
• A document index captures page text, normalized search text, headings, medication-like terms, and short summaries for the live model.
• A transcript parser reacts to the assistant’s output: jump to page (digits or words), highlight spans (quotes, headings, med-style patterns).

• Model: gemini-3.1-flash-live-preview (useGeminiLiveDocument.ts).
• Audio: mic PCM at 16 kHz; playback via Web Audio (user gesture unlock).
• Context: up to 60,000 characters of extracted text plus a document map in the system instruction; if extraction is weak, the first page may be sent as a JPEG. Raw PDF bytes are not uploaded wholesale.
• PHI handling: text sent to Gemini is processed through PHI redaction (phiRedaction.ts); the UI can show a count of redacted items before sending.
• Controls: mute/unmute, stop session, ASL mode (see below).

• Runs when document text is available; uses Groq or Gemini per env (extractCriticalActions.ts).
• Gemini: gemini-2.0-flash · Groq: llama-3.3-70b-versatile.

• Uses document text plus heard / reply transcript when available (extractMedicationSchedule.ts).
• PDF via jsPDF + jspdf-autotable · ICS via a small builder.
• Provider: NEXT_PUBLIC_MEDICATION_EXTRACT_PROVIDER — auto groq gemini (auto prefers Groq when its key is set).

• ZIP validation, then opens Google Maps via googleMapsPharmaciesNearZip in lib/maps/externalMapLinks.ts (search query pharmacies near {zip}).

• MediaPipe hand landmarks + letter classification; optional calibration under public/asl/.
• In ASL mode the mic is off; spelled text can be sent into the live session.

• Hero, Upload PDF, and a disabled View demo (coming soon).
• Features section: Framer Motion scroll narrative, or a static list when reduced motion is preferred.
• Theme: light/dark toggle, localStorage, and a beforeInteractive script to reduce theme flash.
• Branding: UI wordmark uses DM Serif Display; body UI uses Geist / Geist Mono (layout.tsx).

Pinned dependency versions (see package.json):

This describes what the code does, not legal advice.

• The PDF is loaded in the browser (blob URL). This project has no server/database that stores uploads.
• Text extraction and rendering use PDF.js on the client.
• Closing the tab clears the in-memory session for this app.

For text that powers the AI features, Navis applies client-side redaction first, then sends the redacted content to Google and/or Groq. The pipeline is intentional:

1. Extract — Full document text is read locally from the PDF (pdfExtract.ts).
2. Redact — That string is passed through redactPHI() in phiRedaction.ts: pattern-based replacements run in the browser before any network call uses the document body.
3. Use — The redacted text becomes documentExtractedText for extractions (critical steps, medications) and for the Gemini Live session context (together with a document map whose summaries are also passed through redaction where applicable).

So for structured text, the default posture is: strip identifiable and sensitive literals first, then transmit—not the other way around.

What the redactor targets (high level). The implementation uses explicit regex categories such as patient and contact names, physician names, MRN / encounter IDs, DOB, SSN (including last-four style lines), street addresses, phone numbers, insurance / member / NPI / DEA style identifiers, encounter and signature timestamps, and patient-portal line references. Matches are replaced with neutral placeholders (e.g. [PHONE REDACTED]), and the UI can surface a count (and category breakdown) of redactions before content is sent to Gemini.

Why this matters. Third-party APIs should receive clinical narrative (medications, instructions, follow-up) with fewer direct identifiers attached. That reduces accidental exposure of names, numbers, and addresses in the payload when the patterns match.

Limits you should treat seriously.

• Pattern-based, not perfect. Redaction is regex-driven. Novel name spellings, uncommon formats, OCR quirks, or identifiers that do not match a pattern can still appear in outgoing text. This is not a guarantee of de-identification under HIPAA or other rules.
• Vision fallback. If extracted text is weak, a first-page image may be sent for Gemini Live. That image is not run through the same text redaction pipeline—anything visible on the page (including names in a header) could still appear. Prefer text-first PDFs for sensitive demos.
• Voice and ASL. Speech audio and camera / spelled input are not “redacted” like document text; they are processed as audio or text in the live session under Google’s terms.
• Medication prompts may include heard/reply transcript strings; those reflect what was said in session, not a full re-redaction of the PDF.

• Node.js (LTS) and npm (or compatible package manager).
• A modern browser with Web Audio, **getUserMedia** (mic), and optionally camera (ASL).

Copy .env.example → .env.local. Keys are **NEXT_PUBLIC_*** (exposed to the client)—restrict by HTTP referrer in Google AI Studio / Groq.

The marketing shell works without keys; AI features need keys as above.

npm install
cp .env.example .env.local
# Edit .env.local — set NEXT_PUBLIC_GOOGLE_API_KEY at minimum for voice chat
npm run dev

Open http://localhost:3000, upload a PDF, and use the document workspace.

1. Push the repo to GitHub, GitLab, or Bitbucket.
2. Import the project at vercel.com; use the Next.js preset and default install/build commands.
3. Add the same **NEXT_PUBLIC_*** variables under Settings → Environment Variables.
4. To limit API usage to production traffic only, attach keys to the Production environment and omit Preview (preview builds will not have working AI until you add keys there).

src/
  app/                 # App Router — routes, layout, global CSS
  components/          # UI: viewer, panels, header, providers, features
  hooks/               # useGeminiLiveDocument (Gemini Live)
  lib/
    asl/               # Hand landmarks, letter classification
    criticalActions/   # Critical steps extraction
    ics/               # .ics builder
    maps/              # Google Maps search URLs
    medications/       # Medication extract, PDF export
    geminiLiveAudio.ts
    documentIndex.ts
    pdfExtract.ts
    transcriptParser.ts
    phiRedaction.ts
public/
  asl/                 # Optional ASL calibration JSON
  logo.svg

• Use restricted browser keys (referrer / HTTP restrictions).
• Never commit production keys; NEXT_PUBLIC_* is visible in the client bundle by design.

Navis is a tool to organize and read information. It does not provide medical advice, diagnosis, or treatment. Always follow your clinician and verify medications with a licensed professional or pharmacist.

• Browser extension — Coming soon: work with documents from the web without leaving your workflow.
• Multilingual UI — Broader language controls aligned with product copy.

Private project ("private": true in package.json). Adjust for your team as needed.

Navis · Built with Next.js

npm package name: recast